Neural Networks for EMC Modeling of Airplanes Metz

Neural Networks for EMC Modeling of Airplanes Metz, 3. 7. 2010 Vlastimil Koudelka Department of Radio Electronics FEKT BUT xkoude 08@stud. feec. vutbr. cz

Outline q Introduction to artificial neural networks (ANNs) q Neural networks abilities q Evaluation of EM immunity of layered str. q Application of neural classifier in EMC q Regression neural network based optimization q Related problems q Conclusion 2 xkoude 08@stud. feec. vutbr. cz

Introduction to ANNs (1) q Highly parallel structures q Basic element: Neuron q Organized to the layers q Adaptive nonlinear mapping (learning) q Nonlinear separable classification q Optimization features 3 xkoude 08@stud. feec. vutbr. cz (2) 1, 2 (2) 2, 2 (2) 3, 2 W W W n S 2) b 2 ( f(n)

Introduction to ANNs (2) Regression 4 Classification q Multi layered perceptron (MLP) q MLP q Radial basis network (RBF) q Probabilistic NN (PNN) q General regression network (GRN) q Self organizing map (SOM) xkoude 08@stud. feec. vutbr. cz

Introduction to ANNs (3) Optimization q Self organizing map q Self adopted GRNN q Hopfield neural network 5 xkoude 08@stud. feec. vutbr. cz

NN applications: objectives and motivations q Behavioral modeling: continuous models, computational efficiency q Neural models: composite materials, equipment input impedances, aircraft fuel gauge, field levels q Pre-processing, post-processing and optimization tools q Offline training / Online responses q Suitable for direct parallel implementation q Noise suppression (Bayesian regularization) q Multidimensionality, adaptability, generalization, robustness 6 xkoude 08@stud. feec. vutbr. cz

Evaluation of EM immunity of layered str. (1) q At three virtual probes the electromagnetic fields values are estimated by ANN q Harmonic and pulsed wave excitations (Gaussian pulse) q MLP, RBF, PNN 7 xkoude 08@stud. feec. vutbr. cz

Evaluation of EM immunity of layered str. (2) Harmonic waves Pulsed waves q Dependencies of EM field values on the electrical parameters q Pulsed waves: the electric field intensity is expressed in its effective values to respect the mean stress of a virtual device. q MLP, RBF, PNN 8 xkoude 08@stud. feec. vutbr. cz

Application of neural classifier in EMC Probability density function of exceeding prescribed limit q Classification of structures with various electrical parameters q Probability of EM structure successfulness is estimated by ANN q Probabilistic neural network 9 xkoude 08@stud. feec. vutbr. cz

Optimization example 10 1) Initial set of trial solutions consisting of n+1 samples 2) Founded minima of the regression surface is taken as a new training pattern 3) After several iterations the GRNN is well adapted to unknown function f (x) xkoude 08@stud. feec. vutbr. cz

Related problems q Black box modeling (interpretation of NN weights and biases) q Validation techniques: cross validation NN error estimation (perturbation analyses) q Training set compilation: preprocessing and initialization q A number of neurons: clustering problem, regularization q Efficiency and performance of training algorithms: Benchmarks q Stability and robustness of dynamical NNs 11 xkoude 08@stud. feec. vutbr. cz

Conclusion q Regression, classification, optimization q Computational efficiency, good generalization abilities q Pre-processing, post-processing and optimization tools q Wide area of applications (neural network adaptation) q Shielding efficiency, EM structure classification, material modeling q Black box modeling, validation, benchmarking 12 xkoude 08@stud. feec. vutbr. cz

Contact Vlastimil Koudelka xkoude 08@stud. feec. vutbr. cz Department of Radio Electronics, Brno University of Technology Purkynova 118, 612 00 Brno, Czech Republic Tel: +420 541 149 117 Fax: +420 541 149 244 13 xkoude 08@stud. feec. vutbr. cz

This work was supported by the project CZ. 1. 07/2. 3. 00/09. 0092 Communication Systems for Emerging Frequency Bands 14 xkoude 08@stud. feec. vutbr. cz